A remote sensing estimation method for carbon emissions based on nighttime light images

Abstract

The present invention discloses a carbon emission remote sensing estimation method based on night light images, comprising: acquiring night light images and preprocessing the night light images; roughly dividing the preprocessed night light images by a clustering method to obtain a target area, segmenting the target area to obtain regional distribution maps of different levels; processing the night light images and regional distribution maps of different levels by an AI model to generate a carbon emission distribution map.

Description

Technical Field

[0001] The present invention belongs to the technical field of carbon emission estimation, and in particular relates to a carbon emission remote sensing estimation method based on nighttime light images. Background Art

[0002] Traditional carbon emission estimation methods mainly rely on energy consumption statistics and emission factor methods. However, these methods have many problems in data acquisition, statistical errors and update cycles. Therefore, it is of great significance to find a more efficient and accurate carbon emission estimation method.

[0003] With the rapid development of remote sensing technology, it has shown new possibilities in carbon emission estimation. Remote sensing technology obtains surface information through sensors such as satellites or aircraft, and has the advantages of wide coverage, large amount of information, and short update cycle. In particular, night light images, as an important data source of remote sensing technology, can reflect the spatial distribution and intensity of human activities, and are widely used in the research of urbanization, economic development, and energy consumption.

[0004] Night light images refer to the nighttime surface light information obtained by sensors such as satellites or aircraft. This information mainly comes from lighting facilities for human activities, such as street lights, residential lights, and industrial lights. By analyzing the brightness, distribution, and changes of night light images, the spatial distribution and intensity of human activities can be reflected. Then, a correlation is established with energy consumption and carbon emissions. Specifically, the brightness of night light images is positively correlated with energy consumption. Under the same other conditions, the higher the light brightness, the greater the energy consumption in the area, which may lead to higher carbon emissions. In addition, the distribution and changes of night light images can also reflect macro factors such as urbanization process and economic development trends, which will also have an impact on carbon emissions.

[0005] As research deepens, remote sensing estimation methods for carbon emissions have evolved from simple statistical analysis to complex estimation model construction. Early studies were mainly based on simple statistical analysis methods, which established simple linear or nonlinear relationship models by comparing nighttime light brightness and carbon emissions data in different regions. However, such models have limited fitting capabilities and cannot effectively estimate carbon emissions, and the estimation results often have large errors. At present, there is a need to pursue more accurate and stable remote sensing estimation methods in order to provide more accurate and comprehensive carbon emissions monitoring and regulation. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a carbon emission remote sensing estimation method based on night light images to solve the problems existing in the above prior art.

[0007] To achieve the above object, the present invention provides a carbon emission remote sensing estimation method based on nighttime light images, comprising:

[0008] Acquire night light images and pre-process the night light images;

[0009] The pre-processed night light image is roughly divided by clustering method to obtain the target area, and the target area is segmented to obtain regional distribution maps of different levels;

[0010] The AI ​​model is used to process night light images and regional distribution maps of different levels to generate a carbon emissions distribution map.

[0011] Optionally, the night light image adopts the NPP-VIIRS night light remote sensing dataset.

[0012] Optionally, the process of preprocessing the night light image includes:

[0013] The coordinates of the night light images are identical and masked, and the night light images after the masking are subjected to data correction, wherein the data correction process includes: establishing a correction model through a power equation to mutually correct the night light images after the masking; for the night light images collected by different sensors, the mean of the mutually corrected night light images is calculated to correct different sensors and generate the first night light image; for the night light images in different years, if the pixel DN values ​​of the first night light images of the current year and the next year are both greater than the first night light image of the current year, the pixel DN value of the first night light image of the current year is replaced by the pixel DN value of the first night light image of the previous year to perform corrections for different years.

[0014] Optionally, the preprocessed nighttime light image is roughly divided by a K-means algorithm, wherein the optimal number of clusters of the K-means algorithm is determined according to an elbow method.

[0015] Optionally, the process of obtaining the optimal number of clusters includes:

[0016] Initially, the number of clusters k=2 is set, and clustering calculation is performed on the pixels of the preprocessed night light image, and the total number of pixels Z in the cluster corresponding to the initial target area is counted;

[0017] The number of clusters is gradually increased by 1, and the clustering results corresponding to the k value in the gradually increasing process are counted, and the total number of pixels Y under the corresponding cluster of the target area is counted;

[0018] Calculate the pixel drop rate p = Y / Z containing the crack cluster;

[0019] Based on the iterative calculation of all pixel drop rates p, the elbow method is used to calculate the inflection point. When it is calculated that p has an inflection point, the iteration is stopped. Otherwise, the above steps of gradually increasing the number of clusters by 1, calculating the pixel drop rate, and using the elbow method to calculate the inflection point are repeated to obtain the number of clusters corresponding to the inflection point as the optimal number of clusters.

[0020] Optionally, the target area is segmented through a fully convolutional neural network to obtain regional distribution maps of different levels, wherein the regional distribution maps of different levels include light emission areas, slightly heavy emission areas, relatively heavy emission areas, heavy emission areas and relatively heavy emission areas, and the pixels in the light emission areas, slightly heavy emission areas, relatively heavy emission areas, heavy emission areas and relatively heavy emission areas are respectively labeled differently.

[0021] Optionally, the fully convolutional neural network is optimized through data migration and fine-tuning methods.

[0022] Optionally, the fine-tuning method includes adjusting network parameters, selectively freezing layers, and adjusting learning rates.

[0023] Optionally, the AI ​​model includes an input layer, a feature encoder, and a Softmax classifier output layer; the output end of the input layer is cascaded with the input end of the feature encoder, and the output end of the feature encoder is cascaded with the input layer of the Softmax classifier;

[0024] The feature encoder includes a first convolutional layer, a first normalization layer, a first ReLU activation layer, a second convolutional layer, a second normalization layer, a second ReLU activation layer, a third convolutional layer, a third normalization layer, and a third ReLU activation layer, which are sequentially cascaded; the Softmax classifier includes a global average pooling layer, a fully connected layer, and a Softmax layer.

[0025] Optionally, the AI ​​model is trained using a cross loss function.

[0026] Compared with the prior art, the present invention has the following advantages and technical effects:

[0027] Through the above technical solution, the present invention can track and analyze the continuity and dynamics of light changes, providing a basis for long-term monitoring and trend analysis of carbon emissions. At the same time, time series data can provide more comprehensive information and reduce accidental errors that may be caused by single data acquisition. Night light image preprocessing can significantly improve the quality and accuracy of night light images through preprocessing operations such as radiation calibration, provide a more reliable data source for subsequent analysis, and thus improve the accuracy of carbon emission estimates. Using clustering algorithms to perform coarse classification of night light images can quickly identify areas with denser lights, providing a basis for subsequent fine classification. Fine classification can further divide densely lit areas into different categories and levels, providing auxiliary data for subsequent carbon emission distribution. Through sample learning and training through AI models, the correlation between carbon emissions and image data can be quickly and more efficiently fitted, improving the efficiency and accuracy of carbon emission estimates. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0029] Figure 1 It is a schematic diagram of the overall method flow of an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of a preprocessing method according to an embodiment of the present invention;

[0031] Figure 3 Schematic diagram of a data correction method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0033] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0034] Embodiment 1

[0035] like Figure 1As shown, a carbon emission remote sensing estimation method based on night light images is provided in this embodiment, and the method mainly includes the following contents: obtaining time series data of night light images, and performing relevant preprocessing on the night light images to perform image correction and denoising. After preprocessing, the area is divided according to the land features presented in the night light images, and the time series characteristics of the night light impact are combined with the corresponding carbon emissions, and the corresponding connection is fitted through the relevant AI model. The night light impact at the current time is input into the AI ​​model to obtain the corresponding carbon emissions.

[0036] As some real-time examples, the present invention uses the NPP-VIIRS night light remote sensing dataset, using the VIIRS sensor to provide nighttime visible light data for the ocean and atmosphere. In the highest resolution mode of VIIRS, the ground resolution is usually around 100 meters. The dataset has the following characteristics: high data resolution: The NPP-VIIRS night light remote sensing dataset provides the radiance value of each pixel, which means that the dataset has a high spatial resolution and can accurately record the radiation intensity of lights on the ground. Global coverage: The NPP-VIIRS night light remote sensing dataset has global coverage capabilities and can obtain night light conditions around the world. It can provide a wider range for related research. Time series data: The NPP-VIIRS night light remote sensing dataset can obtain continuous observation data for monitoring time changes. This enables relevant researchers to analyze the changing trends of urban lights over long time scales. The present invention can effectively estimate and measure carbon emissions based on the dataset through the advantages provided by the above dataset.

[0037] As some embodiments, the present invention uses relevant preprocessing methods to perform image-by-image correction on the above-mentioned nighttime light images, calibrates the image data, and ensures the accuracy and credibility of satellite observation data. Specifically, calibration is the process of converting the brightness grayscale value of an image into absolute radiant brightness. Masking is performed on areas of the image data that do not need to be measured, and the image data is clarified and unified through the above-mentioned relevant data preprocessing methods. At the same time, other preprocessing methods can be used to further enhance, denoise, or uniformly calibrate the image.

[0038] As some embodiments, the night light images are initially classified by clustering methods to classify areas where the lights are more concentrated or the corresponding carbon emissions are more intensive. The areas correspond to commercial areas, residential areas, industrial areas and other areas with relatively strong carbon emissions. The clustering method selects the k-means clustering algorithm, which is an unsupervised learning algorithm for dividing data into K clusters. During the process, each data point is iteratively assigned to the nearest cluster center, and then the center of each cluster is recalculated. This process is repeated until a certain stop condition is met. Its advantages include simplicity, speed and ease of implementation. After the initial classification is completed, the corresponding relevant deep learning model is used to refine the division of the light-intensive or carbon-intensive areas. Relying on the powerful feature learning and pattern recognition capabilities of the deep learning model, the divided areas are further classified in the case of the initial night light image classification. After the classification is completed, the areas outside the initial classification area are regarded as light areas, and the refined areas are determined according to different levels.

[0039] As some embodiments, relevant carbon emissions are obtained, and carbon emissions can be monitored and obtained through statistics from environmental protection departments or relevant agencies. These agencies will regularly publish reports on carbon emissions, including carbon emission data in the fields of industry, transportation, energy, etc., such as the database of the Ministry of Ecology and Environment, the greenhouse gas emission coefficient library of China's product life cycle, etc. Or obtain it through research reports from academic research institutions and universities. Many academic research institutions and universities conduct research related to the environment and climate change, and publicly publish their research results. By searching the websites or academic databases of these institutions, you can find relevant carbon emission data and research reports.

[0040] As some embodiments, after the carbon emission data and the divided night light images are prepared, the above data are fitted through the AI ​​model. There is a certain correlation between the DN value of the pixel in the night light image and the carbon emission. The existing relevant researchers fit the corresponding relationship through the relevant first or second regression model. The fitting ability of the regression model is limited and it cannot effectively estimate it. The present invention uses the AI ​​model to fit the relationship between its pixel DN value and carbon emission data, and accelerates the training process of its AI model through relevant pre-training and fine-tuning methods, so as to effectively estimate the carbon emission data.

[0041] The above-mentioned related contents are elaborated in detail with reference to the relevant drawings.

[0042] The present invention provides a carbon emission remote sensing estimation method based on night light images. The light image data is clarified by preprocessing related methods to provide an effective data basis. The relevant areas are divided into different levels by coarse classification and fine classification methods. Based on the different levels, the relationship between carbon emissions and night light image data is fitted by an AI model, so that the carbon emission data can be effectively estimated.

[0043] like Figure 1 As shown, the above method specifically comprises the following steps:

[0044] 1) Obtaining time series data of night light images, and preprocessing the night light images to calibrate the night light images and improve the quality of the night light images;

[0045] 2) According to the different regional characteristics shown by night light images, we first use clustering algorithms to roughly classify them, divide the areas with denser lights, and then finely classify the areas with denser lights, and mark the categories and levels of different areas on the classified images;

[0046] 3) Obtain relevant carbon emission data, combine the categories and levels of different areas in the image, conduct sample learning and training on the AI ​​model to fit the correlation between carbon emissions and image data, and obtain the current night light image. Based on the light image at the current time, the AI ​​model recognizes and processes the output and obtains the corresponding carbon emission data.

[0047] The above steps are described in detail:

[0048] In step 1), the night light image data uses the NPP-VIIRS night light remote sensing dataset, and preprocesses the night light image data after collection, such as Figure 2 As shown, the night light image data preprocessing process includes: coordinate unification, mask processing and data calibration.

[0049] like Figure 2 As shown, the coordinates are unified in projection conversion and mask processing. By using the [Data Management Tools] [Projection and Transformation] [Projection] in ArcGIS, the projection of each image is converted to the Albers projection Asia_North_Albers_Equal_Area_Conic (equal area conic projection), and the central meridian and standard parallel related parameters are determined according to their corresponding areas. The night light remote sensing data is clipped using the national boundary vector data mask using the [Spatial Analysis Tools] [Extraction Analysis] [Extraction by Mask] in ArcGIS, and the resolution of the night light remote sensing data is resampled to be consistent with the carbon emission data distribution data through the [Data Management Tools] [Raster] [Raster Processing] [Resampling] in ArcGIS.

[0050] like Figure 3 As shown, in the image data correction, the reference image is used to correct the image to be corrected. The reference image is a high-quality image generated by relevant artificial screening or data simulation in the night light image, which is used to provide the basis and standard for correction of the image to be corrected. The image to be corrected is a night light image with poor quality that needs to be processed. It includes mutual correction and saturation correction. After determining the image to be corrected and the reference image, the correlation between the image to be corrected and the reference image fitted by the power equation is the best. Therefore, the power equation is selected to establish the correction model for mutual correction and saturation correction.

[0051] DN cal =a×DN b ;

[0052] Among them, DN is the pixel DN value in the image to be corrected, DN cal is the DN value of the pixel of the corrected image, a and b are different parameters corresponding to the power regression, and the images to be corrected and the reference images of several periods are subjected to power regression analysis to obtain different parameters corresponding to the power regression. The images in the long-term series of night light image data sets after mutual correction are comparable, and the saturation of the pixel DN value is weakened in each period of image to achieve mutual correction and saturation correction.

[0053] After mutual correction and saturation correction, there are differences between the images of the same year acquired by multiple sensors because different satellite sensors have differences and sensors are affected by various factors when acquiring images. Although the mutual correction of the invariant target area method reduces this difference, it does not completely eliminate it. The two images of the same year still exist at the same time. In order to make full use of the night light images of the same year acquired independently by multiple sensors and solve the discontinuity problem between images in the mutually corrected image dataset, the mutually corrected image dataset is corrected between images of the same year acquired by different sensors.

[0054]

[0055] in, are the DN values ​​of pixel i in the night light images obtained by two different sensors after mutual correction in the nth year, and DN (n，i) It represents the DN value of pixel i in the image of year n after correction.

[0056] Since the mutual correction of image data sets and the correction between images of the same year obtained by different sensors have not solved the problem of abnormal fluctuations in pixel DN values ​​between images of different years obtained by multiple sensors, it is necessary to perform correction between image data of different years obtained by multiple sensors. The correction basis is as follows: (1) When the pixel DN value in the image of the next year is equal to 0, the pixel DN at the same position in the image of the previous year should also be equal to 0; (2) When the pixel DN value in the image of the next year is not equal to 0, the pixel DN value in the image of the previous year should not be greater than the pixel DN value at the same position in the image of the next year. The pixel correction between images of different years is:

[0057]

[0058] Among them, DN (n-1，i) , D.N. (n，i) , D.N. (n+1，i) They represent the DN value of pixel i in the night light images of the n-1th year, the nth year and the n+1th year after mutual correction and the correction between the images of the same year acquired by multiple sensors.

[0059] Through the above content, the night light images are calibrated with each other, different sensors in the same year are calibrated, and corrections in different years are performed. This method solves the problem of instability and discontinuity between images in the long-term series of night light image datasets, and to a certain extent weakens the saturation of the DN value of the image pixels in the image dataset.

[0060] In step 2), the k-means clustering algorithm is first used to perform a rough classification. During the rough classification process, the k-means clustering algorithm is used to classify the areas with denser lights, which is achieved by removing the pixels in the areas with denser lights.

[0061] The traditional K-means clustering method is a very popular and widely used unsupervised learning algorithm for dividing the points in a data set into K different clusters. The following are the basic steps of the K-means clustering method: Initialization: Randomly select K points as the initial centroids (or cluster centers). Iteration: For each point in the data set, calculate its distance to the K centroids (such as using Euclidean distance). Assign each point to the cluster corresponding to the nearest centroid. For each cluster, recalculate its centroid as the mean of all points in the cluster. Repeat the above steps until the position of the centroid no longer changes significantly (or the preset maximum number of iterations is reached). End: Output the final cluster and the corresponding centroid.

[0062] In night light images, in addition to being densely lit, areas with denser lights also have higher pixel DN values ​​than areas with sparser lights. In night light images, areas with denser lights will be divided into clusters with higher pixel DN values ​​after clustering. However, clusters with higher pixel DN values ​​contain not only pixels from areas with denser lights, but also pixels from outside the areas that are close to the pixels from areas with denser lights. To solve this problem, the present invention uses the k-means algorithm to gradually eliminate pixels from outside the areas in clusters corresponding to areas with denser lights, so as to effectively divide areas with denser lights.

[0063] In nighttime light images, the K-means clustering algorithm is used. However, directly setting a small K value such as 2 may cause pixels with denser lights to be mixed with pixels with sparser lights in the same cluster, because the pixels around the denser light areas may have similar pixel DN values ​​to the denser light pixels.

[0064] In order to more accurately separate crack pixels, the present invention can gradually increase the number of clusters (K value). As the K value increases, pixels that are not in the area and were originally mixed in a cluster with a higher pixel DN value may be assigned to other clusters, thereby making the pixels in the area with denser lights more concentrated.

[0065] In order to avoid the over-clustering problem (i.e., the K value is set too large, resulting in the weakening of the crack characteristics), the present invention needs to calculate an optimal K value. This can usually be achieved by the following methods:

[0066] Elbow method: Draw a curve graph of the number of pixels and K value in the cluster of densely lit areas in the cluster results. In this cluster, when the K value gradually increases from a small value, the number of pixels will decrease rapidly; and when the K value continues to increase, the rate of decrease of the number of pixels will slow down, forming an inflection point similar to an "elbow". The K value corresponding to this inflection point is usually considered to be a better K value.

[0067] Silhouette coefficient: For each sample in the data set, calculate its silhouette coefficient, which measures the closeness of the sample to other samples in its cluster and the degree of separation from other clusters. Then calculate the average silhouette coefficient of all samples and find the K value that maximizes the average silhouette coefficient.

[0068] In the present invention, classification is performed by clustering algorithm based on the elbow method. Other methods can be used to calculate the average silhouette coefficient and CH index under different K values, and find the K value that makes these indicators optimal, but no more details are given. The present invention determines the optimal K value by the elbow method. First, a series of different K values ​​are set (such as gradually increasing from 2), and the K-means algorithm is used to cluster the pixels in the night light image. Then, a curve graph of the number of pixels and the K value is drawn, and it is observed whether there is an obvious "elbow" inflection point.

[0069] In the clustering process with different k values, as the number of clusters gradually increases, the distance between the non-target area pixels in the cluster containing the pixels in the area with denser lights (target area) and the cluster center gradually shortens, which indicates that the non-target area pixels are getting closer or more similar to the target area pixels in performance. The number of non-target area pixels in the entire cluster gradually decreases from the boundary to the center of the cluster corresponding to the target area, indicating that the non-target area pixels are separated from the cluster corresponding to the target area and reallocated to other clusters. Based on the above phenomenon, in the above process, the number of pixels in the non-target area in the cluster containing the target area increases with each increase in the k value for clustering, and the number of pixels in the cluster corresponding to the target area that are divided into other clusters shows a gradient decrease trend and has convergence characteristics.

[0070] The present invention calculates the optimal number of clusters by taking the ratio p of the total number of pixels of the cluster containing the target area pixels corresponding to each set k value and the total number of pixels of the cluster containing the target area pixels in the initial k=2 clusters as the drop rate of the target area pixels. The inflection point, i.e., the position of the optimal number of clusters, is determined by the elbow method.

[0071] The specific calculation process of the optimal number of clusters is as follows: first, set the initial k value, and use the initial k value to calculate the total number of pixels under the initial target area cluster. After the initial k value is set and clustering is completed, the k value is gradually increased by 1, and the clustering results corresponding to the k value in the gradual increase process are counted, and the total number of pixels under the target area cluster is counted; then, calculate the ratio p of the total number of pixels in the cluster containing the target area pixels after the k value is gradually increased to the total number of pixels in the cluster containing the target area pixels when the initial k=2 clustering is performed; then, draw a curve of p changing with the number of clusters, and find the point where the rate of change of the p value is significantly reduced (i.e., the inflection point), and the number of clusters corresponding to this point is the optimal number of clusters.

[0072] The specific process is as follows:

[0073] 1. Set the number of clusters k = 2, perform clustering calculation on the pixels of the night light image, and count the total number of pixels Z under the initial target area cluster;

[0074] 2. The k value is gradually increased by 1, and the clustering results corresponding to the k value in the gradual increase process are counted, and the total number of pixels Y under the corresponding cluster of the target area is counted;

[0075] 3. Calculate the pixel drop rate p = Y / Z containing the crack cluster;

[0076] 4. Based on the iterative calculation of all pixel drop rates p, use the elbow method to calculate the inflection point. When the inflection point is calculated, stop the iteration, otherwise repeat steps 2, 3, and 4. Output the optimal number of clusters corresponding to the corresponding inflection point.

[0077] In order to facilitate rapid calculation and processing in night light images, several images are extracted from the data set of night light images as samples for obtaining the above-mentioned optimal clustering number. Under the samples, the corresponding optimal clustering number is obtained through the above-mentioned technical solution. When there are multiple optimal clustering numbers, they can be determined by the mean or the value with the largest number of occurrences. The night light images are pixel clustered according to the optimal clustering number, and the clustered pixels are divided into target areas along the edge pixels in the clustered pixels. The night light images are all images in the same area, and can be roughly clustered by the same optimal clustering number, without the need to process the above-mentioned images one by one.

[0078] After the coarse classification is completed, the target area generated by the above clustering algorithm, that is, the initial classification area, is refined through the deep learning model. The coarse division is to distinguish between areas with more concentrated lights and areas with sparse lights. Areas with more concentrated lights include areas with more carbon emissions, such as commercial areas, residential areas, and industrial areas. Sparsely lit areas include areas with less carbon emissions, such as rural areas and fields. The correlation between the DN value of pixels in the night light image and carbon emissions will also produce certain areas according to the different regions in which they are located. Therefore, in order to accurately estimate carbon emissions, regional differentiation is also an indispensable link.

[0079] The present invention performs refinement classification on areas where light is more concentrated. The input data is to fill the DN value of pixels in non-target areas with 0, and the DN value remains unchanged. The refinement is mainly to distinguish areas at different levels. Based on carbon emissions, the areas are divided into light emission areas, heavier emission areas, severe emission areas and heavier emission areas. The non-target areas in the coarse classification are regarded as light emission areas, and the pixels corresponding to the areas of different levels are numbered separately.

[0080] In the subsequent segmentation of heavy emission areas of different levels, the present invention performs segmentation through a related semantic segmentation model, and the semantic segmentation model adopts a fully convolutional neural network FCN network model.

[0081] The FCN network structure is mainly divided into two parts: the full convolution part and the deconvolution part.

[0082] Full convolution part: Use the classic convolutional neural network structure (such as ResNet) to extract features of the target area of ​​the night light image. This part of the network gradually reduces the size of the feature map through convolution and pooling operations, while increasing the number of channels of the feature map, thereby capturing multi-level information in the image.

[0083] Deconvolution part: The feature map output by the full convolution part is restored to the same size as the original input image through upsampling operations (such as transposed convolution). During the upsampling process, new pixel values ​​can be generated by interpolation or learning to restore the spatial resolution of the image. In FCN, deconvolution layers are usually used to implement upsampling operations.

[0084] In order to obtain more refined segmentation results, FCN uses skip connections to fuse feature maps at different levels. Specifically, the shallow detail information can be combined with the deep semantic information to improve the accuracy and robustness of the segmentation.

[0085] Deconvolution can be seen as the inverse process of convolution, but in actual calculations it is usually implemented by transposing the convolution kernel. The specific calculation formula is similar to convolution, but the convolution kernel needs to be transposed.

[0086] Among them, FCN is a commonly used segmentation model, and its structure can adopt a related fully convolutional network structure, which will not be elaborated here. In order to adapt the fully convolutional neural network to the relevant recognition and segmentation of night light images, the present invention provides an adaptive setting of its model through relevant migration and fine-tuning methods.

[0087] First, the fully convolutional neural network is trained using the relevant data migration method. It can be pre-trained using the relevant source data set: such as the Cityscapes data set. This data set is an existing data set that is easy to obtain, and can be used to train the fully convolutional neural network by including images before segmentation and results after segmentation. This data set contains a wealth of natural scenes and a variety of target categories, and the parameters of the model can be pre-trained. Target data set: For the image segmentation task of night light images, a special night light image data set is prepared through relevant manual annotation. Since the parameters of the fully convolutional neural network have been adjusted during the pre-training process, the scale of the target data set does not need to be particularly large, and it can also be used effectively.

[0088] Use the source dataset to pre-train a model and save the pre-trained model parameters. Directly use the model pair pre-trained by the source dataset as the initial parameters of the model, or choose to use an image segmentation model parameter that has been trained in other applications as the pre-trained model parameter as the initial parameters of the model. After deploying the full convolutional neural network with the initial parameters, fine-tune the model by using the target dataset as the training set.

[0089] The relevant contents of fine-tuning include the following:

[0090] 1. Fine-tune network parameters:

[0091] After deploying the parameters of the model generated after pre-training to the fully convolutional neural network model, these weights are fine-tuned using the dataset of the new task. Fine-tuning is an iterative process that uses optimization algorithms such as gradient descent to update the parameters of the model to minimize the loss function on the new task. Since the pre-trained model already has good initial weights, a smaller learning rate is usually used to ensure the stability of the fine-tuning process and prevent large deviations in the pre-trained parameters.

[0092] The gradient descent optimization process is expressed as:

[0093]

[0094] Among them, θ t+1 and θ t Respectively represent the values ​​of parameters after and before updating, α represents the learning rate, is the gradient of the loss function with respect to the parameters.

[0095] 2. Selectively freeze layers to balance new and old knowledge:

[0096] During fine-tuning, some layers in the pre-trained model are selectively frozen to prevent the weights in these layers from being updated too quickly. Freezing the layers means that the weights of these layers remain unchanged during training, thereby retaining the knowledge learned by the pre-trained model on the source task to balance the new and old knowledge, ensuring that the model can learn new task-specific features while retaining the general feature representation.

[0097] 3.Flexibly adjust the learning rate strategy:

[0098] For the fine-tuned layers, different learning rate strategies are adopted. Usually, we set a small initial learning rate for these layers to ensure a stable learning process. As the training progresses, the learning rate is adjusted according to the performance of the model, such as using a learning rate decay strategy to gradually reduce the learning rate to promote model convergence. In addition, different learning rates are set for different layers to better balance the integration of new and old knowledge.

[0099] Through the above steps, the pre-trained model can be effectively used to accelerate the training process of new tasks and improve the performance of the model on new tasks.

[0100] In step 4), the relevant contents of the AI ​​model of the present invention are as follows:

[0101] Night light image data and regional division data are combined with carbon emission distribution maps as the training set and test set.

[0102] After the above data is prepared, its AI model consists of an input layer, a feature encoder, and a Softmax classifier output layer; the output of the input layer is cascaded with the input of the feature encoder, and the output of the feature encoder is cascaded with the input layer of the Softmax classifier; the feature encoder includes the first convolution layer, the first normalization layer, the first ReLU activation layer, the second convolution layer, the second normalization layer, the second ReLU activation layer, the third convolution layer, the third normalization layer, and the third ReLU activation layer, which are cascaded in sequence; the Softmax classifier includes a global average pooling layer, a fully connected layer, and a Softmax layer.

[0103] In a further embodiment, the parameters of the feature encoder are set as follows;

[0104] The first convolutional layer has a feature map of 32 output feature data, a convolution filter size of 3×3, a padding of 1, and a moving stride of 1; the first normalization layer has a feature map of 32 output feature data; the first ReLU activation layer has a feature map of 32 output feature data; the second convolutional layer has a feature map of 32 output feature data, a convolution filter size of 3×3, a padding of 1, and a moving stride of 1; the second normalization layer has a feature map of 32 output feature data; the second ReLU activation layer has a feature map of 32 output feature data; the third convolutional layer has a feature map of 32 output feature data, a convolution filter size of 3×3, a padding of 1, and a moving stride of 1; the second normalization layer has a feature map of 32 output feature data; the third activation layer has a feature map of 32 output feature data.

[0105] In a further embodiment, the parameters of the Softmax classifier are set as follows:

[0106] The number of feature maps of the output feature data of the global average pooling layer is 32; the number of feature maps of the output feature data of the fully connected layer is 64; the number of maps of the output data of the Softmax layer is 14;

[0107] In a further embodiment, the AI ​​model is trained using the above training set; obtaining the trained AI model specifically includes:

[0108] The weight parameter set of the AI ​​model is initialized using the weight initialization method to initialize the AI ​​model.

[0109] Construct the cross entropy loss function of the AI ​​model; use the training set as the input feature data of the first convolutional layer, the input feature data includes the pixel DN value matrix in the night light image and the label matrix of each area in the area division data, calculate the output feature data of the first convolutional layer, and output it to the i-th output feature map data of the first convolutional layer

[0110]

[0111] Among them, W M1 is the convolution filter matrix in the first convolution layer. The elements in this matrix are the weights in the weight set Θ, and * represents the convolution operation.

[0112] The output feature data of the first convolutional layer is used as the i-th input feature map data of the first normalized layer in the AI ​​model, and the n-th output feature map data of this layer is calculated according to the following formula:

[0113]

[0114] In the formula, The first normalized layer output feature data representing the kth training sample feature, f CBN (·) is the normalization operation.

[0115] The output feature data of the first normalized layer is used as the i-th input feature map data of the first ReLU activation layer in the AI ​​model, and the i-th output feature map data of the layer is calculated according to the following formula:

[0116]

[0117] Among them, f RELU (·) is the rectified linear unit (ReLU) activation function;

[0118] The first RELU activation layer is used as the input feature data of the second convolutional layer, the output feature data of the second convolutional layer is calculated, and output to the i-th output feature map data of the second convolutional layer

[0119]

[0120] Among them, W M2 is the convolution filter matrix in the second convolution layer. The elements in this matrix are the weights in the weight set Θ, and * represents the convolution operation.

[0121] The output feature data of the second convolutional layer is used as the i-th input feature map data of the second normalized layer in the AI ​​model, and the n-th output feature map data of this layer is calculated according to the following formula:

[0122]

[0123] In the formula, The second normalized layer output feature data representing the k-th training sample feature, f CBN (·) is the normalization operation;

[0124] The output feature data of the second normalized layer is used as the i-th input feature map data of the first activation layer in the AI ​​model, and the i-th output feature map data of the layer is calculated according to the following formula:

[0125]

[0126] Among them, f ReLU (·) is the rectified linear unit (ReLU) activation function;

[0127] The second RELU activation layer is used as the input feature data of the second convolutional layer, the output feature data of the third convolutional layer is calculated, and the output feature map data of the third convolutional layer is output to the i-th output feature map data.

[0128]

[0129] Among them, W M3 is the convolution filter matrix in the third convolution layer. The elements in this matrix are the weights in the weight set Θ. * represents the convolution operation.

[0130] The output feature data of the third convolutional layer is used as the i-th input feature map data of the third normalized layer in the AI ​​model, and the i-th output feature map data of this layer is calculated according to the following formula:

[0131]

[0132] In the formula, f i b3 The third normalized layer output feature data representing the i-th training sample feature, f CBN (·) is the normalization operation;

[0133] The output feature data of the third normalized layer is used as the i-th input feature map data of the third activation layer in the AI ​​model, and the i-th output feature map data f of the layer is calculated according to the following formula:i p3 .

[0134] f i p3 =f ReLU (f i b2 )

[0135] Among them, f ReLU (·) is the rectified linear unit (ReLU) activation function;

[0136] The output feature data of the third ReLU activation layer is used as the i-th input feature map data of the global average pooling layer in the Softmax classifier, and the i-th output feature map data of this layer is calculated according to the following formula:

[0137]

[0138] Among them, f CGAP (·) is the global average pooling operation.

[0139] The output feature data of the global average pooling layer is used as the i-th input feature map data of the fully connected layer in the Softmax classifier, and the i-th output feature map data of the layer is calculated according to the following formula:

[0140]

[0141] Among them, f CFC (·) is the fully connected layer operation.

[0142] The output feature data of the fully connected layer is used as the i-th input feature map data of the Softmax layer in the Softmax classifier, and the i-th output feature map data p of the layer is calculated according to the following formula: i .

[0143]

[0144] Among them, p k The output prediction of the output layer of the Softmax classifier representing the characteristics of the i-th training sample is the distribution matrix corresponding to the carbon emissions, and the position of its matrix elements corresponds to the pixel position in the carbon emissions distribution map, f CSoftmax (·) is the Softmax function.

[0145] The AI ​​model uses a cross-entropy loss function; the initialized AI model is then trained to obtain the partial derivative of the loss value with respect to each weight parameter in the weight parameter set; the weight parameters are updated by back propagation using the gradient descent method and the partial derivative values ​​to obtain updated weight parameters; an updated weight parameter set is established based on all updated weight parameters, and the cross-entropy loss function is recalculated based on the updated weight set until the loss function converges to obtain an updated weight set.

[0146] In the present invention, the trained AI model is trained through the test set, the night light image to be identified is input into the pre-trained feature encoder, and then input into the classifier, and the carbon emission distribution matrix corresponding to the night light image is output. The position of each element in the matrix corresponds to the position of the carbon emissions, and its value corresponds to the value of the carbon emissions, so as to generate the corresponding carbon emissions distribution map.

[0147] The carbon emission distribution map can be visualized through relevant software combined with GIS maps, and it can also help relevant personnel analyze carbon emission trends or make further decisions.

[0148] Through the above technical solution, the present invention can track and analyze the continuity and dynamics of light changes, providing a basis for long-term monitoring and trend analysis of carbon emissions. At the same time, time series data can provide more comprehensive information and reduce accidental errors that may be caused by single data acquisition. Night light image preprocessing can significantly improve the quality and accuracy of night light images through preprocessing operations such as radiation calibration, provide a more reliable data source for subsequent analysis, and thus improve the accuracy of carbon emission estimates. Using clustering algorithms to perform coarse classification of night light images can quickly identify areas with denser lights, providing a basis for subsequent fine classification. Fine classification can further divide densely lit areas into different categories and levels, providing auxiliary data for subsequent carbon emission distribution. Through sample learning and training through AI models, the correlation between carbon emissions and image data can be quickly and more efficiently fitted, improving the efficiency and accuracy of carbon emission estimates.

[0149] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A carbon emission remote sensing estimation method based on night light images, characterized in that: include: Acquire night light images and pre-process the night light images; The pre-processed night light image is roughly divided by clustering method to obtain the target area, and the target area is segmented to obtain regional distribution maps of different levels; The AI ​​model is used to process nighttime light images and regional distribution maps of different levels to generate a carbon emissions distribution map; The AI ​​model includes an input layer, a feature encoder, and a Softmax classifier output layer; the output end of the input layer is cascaded with the input end of the feature encoder, and the output end of the feature encoder is cascaded with the input layer of the Softmax classifier; The feature encoder includes a first convolution layer, a first normalization layer, a first ReLU activation layer, a second convolution layer, a second normalization layer, a second ReLU activation layer, a third convolution layer, a third normalization layer, and a third ReLU activation layer, which are sequentially cascaded; the Softmax classifier includes a global average pooling layer, a fully connected layer, and a Softmax layer; The AI ​​model is trained by a cross loss function; AI model training specifically includes: Initialize the weight parameter set of the AI ​​model using a weight initialization method to initialize the AI ​​model; Construct the cross entropy loss function of the AI ​​model; use the training set as the input feature data of the first convolutional layer, the input feature data includes the pixel DN value matrix in the night light image and the label matrix of each area in the area division data, calculate the output feature data of the first convolutional layer, and output it to the i-th output feature map data of the first convolutional layer Among them, W M1 is the convolution filter matrix in the first convolutional layer. The elements in this matrix are the weights in the weight set Θ. * represents the convolution operation. The output feature data of the first convolutional layer is used as the i-th input feature map data of the first normalized layer in the AI ​​model, and the n-th output feature map data of this layer is calculated according to the following formula: In the formula, The first normalized layer output feature data representing the kth training sample feature, f CBN (·) is the normalization operation; The output feature data of the first normalized layer is used as the i-th input feature map data of the first ReLU activation layer in the AI ​​model, and the i-th output feature map data of the layer is calculated according to the following formula: Among them, f RELU (·) is the rectified linear unit (ReLU) activation function; The first RELU activation layer is used as the input feature data of the second convolutional layer, the output feature data of the second convolutional layer is calculated, and the output feature map data of the second convolutional layer is output to the i-th output feature map data. Among them, W M2 is the convolution filter matrix in the second convolutional layer. The elements in this matrix are the weights in the weight set Θ. * represents the convolution operation. The output feature data of the second convolutional layer is used as the i-th input feature map data of the second normalized layer in the AI ​​model, and the n-th output feature map data of this layer is calculated according to the following formula: In the formula, The second normalized layer output feature data representing the k-th training sample feature, f CBN (·) is the normalization operation; The output feature data of the second normalized layer is used as the i-th input feature map data of the first activation layer in the AI ​​model, and the i-th output feature map data of the layer is calculated according to the following formula: Among them, f ReLU (·) is the rectified linear unit (ReLU) activation function; The second RELU activation layer is used as the input feature data of the second convolutional layer, the output feature data of the third convolutional layer is calculated, and the output feature map data of the third convolutional layer is output to the i-th output feature map data. Among them, W M3 is the convolution filter matrix in the third convolution layer. The elements in this matrix are the weights in the weight set Θ. * represents the convolution operation. The output feature data of the third convolutional layer is used as the i-th input feature map data of the third normalized layer in the AI ​​model, and the i-th output feature map data of this layer is calculated according to the following formula: In the formula, f i b3 The third normalized layer output feature data representing the i-th training sample feature, f CBN (·) is the normalization operation; The output feature data of the third normalized layer is used as the i-th input feature map data of the third activation layer in the AI ​​model, and the i-th output feature map data f of the layer is calculated according to the following formula: i p3 ; f i p3 =f ReLU (f i b2 ) Among them, f ReLU (·) is the rectified linear unit (ReLU) activation function; The output feature data of the third ReLU activation layer is used as the i-th input feature map data of the global average pooling layer in the Softmax classifier, and the i-th output feature map data of this layer is calculated according to the following formula: Among them, f CGAP (·) is the global average pooling operation; The output feature data of the global average pooling layer is used as the i-th input feature map data of the fully connected layer in the Softmax classifier, and the i-th output feature map data of the layer is calculated according to the following formula: Among them, f CFC (·) is the fully connected layer operation; The output feature data of the fully connected layer is used as the i-th input feature map data of the Softmax layer in the Softmax classifier, and the i-th output feature map data p of the layer is calculated according to the following formula: i ; Among them, p k The output prediction of the output layer of the Softmax classifier representing the characteristics of the i-th training sample is the distribution matrix corresponding to the carbon emissions, and the position of its matrix elements corresponds to the pixel position in the carbon emissions distribution map, f CSoftmax (·) is the Softmax function.

2. The method according to claim 1, characterized in that: The night light image adopts the NPP-VIIRS night light remote sensing dataset.

3. The method according to claim 1, characterized in that: The process of preprocessing night light images includes: The coordinates of the night light images are identical and masked, and data correction is performed on the night light images after masking, wherein the data correction process includes: establishing a correction model through a power equation to mutually correct the night light images after masking; for the night light images collected by different sensors, the mean of the mutually corrected night light images is calculated to correct different sensors and generate the first night light image; for the night light images in different years, if the pixel DN values ​​of the first night light images of the current year and the next year are both greater than the first night light image of the current year, the pixel DN value of the first night light image of the current year is replaced by the pixel DN value of the first night light image of the previous year to perform corrections for different years.

4. The method according to claim 1, characterized in that The preprocessed night light images are roughly divided by the K-means algorithm, where the optimal number of clusters of the K-means algorithm is determined according to the elbow method.

5. The method according to claim 4, characterized in that The process of obtaining the optimal number of clusters includes: Initially, the number of clusters k=2 is set, and clustering calculation is performed on the pixels of the preprocessed night light image, and the total number of pixels Z in the cluster corresponding to the initial target area is counted; The number of clusters is gradually increased by 1, and the clustering results corresponding to the k value in the gradually increasing process are counted, and the total number of pixels Y under the corresponding cluster of the target area is counted; Calculate the pixel drop rate p = Y / Z containing the crack cluster; Based on the iterative calculation of all pixel drop rates p, the elbow method is used to calculate the inflection point. When it is calculated that p has an inflection point, the iteration is stopped. Otherwise, the above steps of gradually increasing the number of clusters by 1, calculating the pixel drop rate, and using the elbow method to calculate the inflection point are repeated to obtain the number of clusters corresponding to the inflection point as the optimal number of clusters.

6. The method according to claim 1, characterized in that The target area is segmented by a fully convolutional neural network to obtain regional distribution maps of different levels, wherein the regional distribution maps of different levels include light emission areas, slightly heavy emission areas, relatively heavy emission areas, heavy emission areas and relatively heavy emission areas, and the pixels in the light emission areas, slightly heavy emission areas, relatively heavy emission areas, heavy emission areas and relatively heavy emission areas are respectively labeled differently.

7. The method according to claim 6, characterized in that The fully convolutional neural network is optimized through data migration and fine-tuning methods.

8. The method according to claim 7, characterized in that The fine-tuning method includes adjusting network parameters, selectively freezing layers, and adjusting learning rates.

Images